A Two Level Monte Carlo Approach To Calculating Expected Value of Sample Information:How to Value a Research Design. Alan Brennan and Jim Chilcott, Samer Kharroubi, Tony O’Hagan University of Sheffield IHEA June 2003 What are EVPI and EVSI? Adoption Decision Choose policy with greatest expected ‘payoff’ Uncertainty Confidence interval e.g. |--------[]---------| implies Adoption Decision could be wrong Perfect Information perfect knowledge about parameter e.g. Value of Information Expected Value of Information [] quantify additional payoff obtained by switching adoption decision after obtaining additional data quantify average additional payoff obtained for range of possible collected data EVPI Expected value of obtaining perfect knowledge EVSI Expected value of obtaining a specific sample size Typical Process Net benefit of alternative policies Illustrative Model a Mean Model Cost of drug % admissions Days in Hospital Cost per Day £ £ % Side effects Change in utility if side effect Duration of side effect (years) £ Total QALY Probabilistic Sensitivity Analysis d e Uncertainty in Parameter Means Standard Deviations Increment Mean Model T0 T1 (T1 over T0) Cost of£drug 1,500 £ 1,000 500 % admissions 10% 8% -2% Days in Hospital6.10 5.20 0.90 Cost 400 per £ Day 400 £ - % Responding 80% 70% Utility change0.3000 if respond 0.3000 Duration of response 3.0 3.0 (years) % Responding Utility change if respond Duration of response (years) Total Cost Illustrative Model b c Number of Patients in the UK 1,000 Threshold cost per QALY £10,000 Number of Patients in the UK Threshold cost per QALY T0 1 2% 1.00 200 Sampled Values T1 1 £ 2% 1.00 200 £ T0 1,000 £ 10% 5.01 589 £ Increment T1 (T1 over T0) 1,499 £ 499 11% 0% 7.05 2.05 589 £ - 10% - 10% 0.1000 0.5 10% 0.0500 1.0 88% 0.3119 3.1 70% 0.2120 4.1 % Side effects 20% 25% -5% Change in utility -0.10 if side effect -0.10 0.00 Duration of side 0.50 effect (years) 0.50 10% 0.02 0.20 5% 0.02 0.20 31% -0.06 0.70 24% -0.14 0.65 - Total Cost 1,208 £ 1,695 Total QALY 0.6175 0.7100 0.0925 Cost per 1,956 £ QALY 2,388 £5,267 £ £ 487 Cost per QALY £ Net Benefit of T1 versus T0 Net Benefit of T1 versus T0 £ 5,405 £437.80 £ 4,967 1,308 £ 0.8394 £ 1,558 £ £7,086 1,937 -18% 0.0998 1.0 -7% -0.08 0.05 £ 629 0.5948 -0.2446 3,256 -£2,570 £4,011 -£3,075 e.g. 1,000 samples … C-E plane Cost Effectiveness Acceptability of T1 versus T0 £2,000 100% CEACs £1,500 £1,000 90% 80% 70% Probability Cost Effective Model Inc Cost £500 £0 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 -£500 60% T1 50% T0 40% 30% -£1,000 20% 10% -£1,500 0% -£2,000 £- £20,000 £40,000 Inc QALY identify single or groups of parameters £80,000 £100,000 £120,000 Threshold (MAICER) EVPI £1,400 Value of Information Partial EVPIs £60,000 £1,200 £1,000 £800 £600 £400 £200 £- All Six Trial + Utility Utility Study Trial on % Response propose specific data collections (n) calculate expected value of different samples EVI for % Responders to T0 Value of Information Partial EVSIs Durations 250 200 150 EVSI (n) EVPI 100 50 0 0 50 100 Sample Size (n) 150 200 £140,000 d NB(d, ) i -i = uncertain model parameters = set of possible decisions = net benefit (λ*QALY – Cost) for decision d, parameters = parameters of interest – possible data collection = other parameters (not of interest, remaining uncertainty) Expected net benefit (1) Baseline decision = max E NB(d, ) d (2) Perfect Information on i = Partial EVSI = (3) – (1) d two expectations Partial EVPI = (2) – (1) (3) Sample Information on i E i max E i NB(d, ) | i = E X max E i NB(d, ) | i d Expected Value of Sample Information 0)Decision model, threshold, priors for uncertain parameters 1) Simulate data collection: • sample parameter(s) of interest once ~ prior • decide on sample size (ni) (1st level) • sample the simulated data | parameter of interest 2) combine prior + simulated data --> simulated posterior 3) now simulate 1000 times parameters of interest ~ simulated posterior unknown parameters ~ prior uncertainty (2nd level) 4) calculate best strategy = highest mean net benefit 5) Loop 1 to 4 say 1,000 times Calculate average net benefits 6) EVSI parameter set = (5) - (mean net benefit | current information) e.g. 1000 * 1000 simulations Bayesian Updating: Normal Prior 0= mean, 0= uncertainty in mean (standard deviation) I 0 1 / 02 = precision of the prior mean 2pop = patient level uncertainty ( needed for update formula) Simulated Data X = sample mean (further data collection e.g. clinical trial ). 2 2pop / n = sample variance I s 1 / 2 = precision of the sample mean . Simulated Posterior N(1, 1 ) 2 /n 2 0 0 s pop 2 0 1 2 1 2 / n 0 s 0 pop Normal Posterior Variance – Implications 2 /n 2 pop 2 0 1 2 2 / n pop 0 . • 1 always < 0 • If n is very small, 1 = almost 0 • If n is very large, 1 = almost 0 Normal Bayesian Update (n=50) Prior Frequency (1000 samples) 300 250 Posterior after sample 1 200 150 Posterior after sample 2 100 Posterior after sample 7 50 0 0% 20% 40% 60% Response Rate (T0) 80% 100% Bayesian Updating: Beta / Binomial • e.g. % responders to a treatment Prior • % responders ~ Beta (a,b) Simulated Data • n cases, • y successful responders Simulated Posterior • % responders ~ Beta (a+y,b+n-y) Bayesian Updating: Gamma / Poisson e.g. no. of side effects a patient experiences in a year Prior side effects per person ~ Gamma (a,b) Simulated Data • n samples, (y1, y2, … yn) • from a Poisson distribution Simulated Posterior • mean side effects per person ~ Gamma (a+ yi , b+n) Bayesian Updating without a Formula • WinBUGS • Put in prior distribution Conjugate Distributions Beta Inverted Beta • Put in data Cauchy 1 Cauchy 2 • MCMC gives posterior Chi (‘000s of iterations) Chi² 1 • Use posterior in model Chi² 2 • Loop to next data sample Erlang Expon. Fisher Gamma • Other approximation methods Inverted Gamma Gumbel Laplace Logistic Lognormal Normal Pareto Power Rayleigh r-Distr. Uniform Student Triangular Weibull EVSI Results for Illustrative Model Value of Information EVI for % Responders to T0 250 200 150 EVSI (n) EVPI 100 50 0 0 50 100 Sample Size (n) 150 200 Common Properties of EVSI curve • Fixed at zero if no sample is collected • Bounded above by EVPI, monotonic, diminishing returns ? EVSI (n) = EVPI * [1 – exp -a*sqrt(n) ] EVI for % Responders to T0 Value of Information EVSI = EVPI * [1- EXP( -0.3813 * sqrt(n) ] 250 200 EVSI (n) 150 EVPI 100 50 Exponential fit 0 0 50 100 150 Sample Size (n) 200 Correct 2 level EVPI Algorithm 0)Decision model, threshold, priors for uncertain parameters 1) Simulate data collection: • sample parameter(s) of interest once ~ prior • decide on sample size (ni) (1st level) • sample the simulated data | parameter of interest 2) combine prior + simulated data --> simulated posterior 3) now simulate 1000 times fixed at sampled value parameters of interest ~ simulated posterior unknown parameters ~ prior uncertainty (2nd level) 4) calculate best strategy = highest mean net benefit 5) Loop 1 to 4 say 1,000 times Calculate average net benefits 6) EVPI parameter set = (5) - (mean net benefit | current information) e.g. 1000 * 1000 simulations Shortcut 1 level EVPI Algorithm 1) Simulate data collection: • sample parameter(s) of interest once ~ prior • decide on sample size (ni) (1st level) • sample the simulated data | parameter of interest 2) fix remaining unknown parameters 2) combine prior + simulated data --> simulated posterior constant at prior mean value 3) now simulate 1000 times E i maxofNB(d, i~| simulated parameters interest i i )posterior d unknown parameters ~ prior uncertainty (2nd level) 4) calculate best strategy = highest mean net benefit 5) Loop 1 to 4 say 1,000 times Calculate average net benefits 6) EVPI parameter set = (5) - (mean net benefit | current information) Accurate if .. (a) net benefit functions are linear functions of the -i for all d,i, and (b) i and -i are independent. How many samples for convergence ? E V P I e s tim a te s - Im p a c t o f M o re S a m p le s 600 550 5 5 0 -6 0 0 500 5 0 0 -5 5 0 450 400 4 5 0 -5 0 0 350 4 0 0 -4 5 0 300 3 5 0 -4 0 0 250 3 0 0 -3 5 0 £334 200 2 5 0 -3 0 0 150 2 0 0 -2 5 0 100 • • 500 750 1000 2000 5000 1000 10000 750 500 100 10 K - o u te r le ve l 10 100 1 5 0 -2 0 0 50 0 1 0 0 -1 5 0 5 0 -1 0 0 0 -5 0 J - in n e r le ve l outer level - over 500 samples, converges to within 1% inner level - required 10,000 samples, to converge to within 2%. How wrong is US 1 level EVPI approach for a non-linear model? • E.g – squared every model parameter • Adjusted R2 for simple linear regression = 0.86 Part b: Comparison of 1 Level US versus 2 Level US in the Two Models 150 100 Linear Model Non-linear Model 50 T r ia l U tility O n ly Difference 0 T r ia l + U tility -5 0 D u r a tio n s -1 0 0 Main Illustrative Model -1 5 0 -2 0 0 -2 5 0 2nd Model (non-linear) Maximum of Monte Carlo Expectations:Upward Bias • simple Monte Carlo estimates are unbiased • But, bias occurs when use maximum of Monte Carlo estimates. E{max(X,Y)} > max{E(X),E(Y)}. E.g. X ~ N(0,1), Y ~ N(0,1) • E{max(X,Y)} = 1.2 , max{E(X),E(Y)} = 1.0 • If variances of X and Y reduces to 0.5, E{max(X,Y)} = 1.08 • E{max(X,Y)} continues to fall variance reduces, but stays > 1.0 1 Partial EVPI = K 1 max d 1toD J k 1 K NBd , J j 1 i j i i k 1 L NB(d, l ) - dmax 1toD L l 1 • Illustrative model:- 1000*1000 not enough to eliminate bias. • Increasing no. of samples reduces bias, since it reduces the variances of the Monte Carlo estimates. Computation Issues: Emulators • Gaussian Processes (Jeremy Oakley, CHEBS) • • • F() ~ NB(d, ) Bayesian non-linear regression (complicated maths) Assumes only a smooth functional form to NB(d, ) Benefits 1. Can emulate complex time consuming models with formula i.e. speed up each sample 2. Can produce a quick approximation to inner expectation for partial EVPI 3. Similar quick approximation for partial EVSI but only for one parameter Summary 1. 2 level algorithm for EVPI and EVSI is correct approach 2. Bayesian Updating for Normal, Beta, Gamma OK Others – WinBUGS / approximations 3. There are issues of computation 4. Shortcut 1 level EVPI algorithm is accurate if … (a) net benefit functions are linear and (b) the parameters are independent. 5. Emulators (e.g. Gaussian Processes) can be helpful